Comment by phaser | Hacker News Reader

Comment by phaser | Hacker News Reader

Here’s how i do it: I create a lot of stuff using AI to the max, but I also spend the necessary of time on reviewing that the AI is producing code that passes my cognitive load standards. this involves some tokens spent on grooming code and documenting well. Most of this is effortless thanks to an AGENTS.md based on this: https://github.com/zakirullin/cognitive-load/blob/main/READM... but i have a good sense of catching when things are getting weird and i steer back.

Then, when credits run out. It’s show time! The code is neatly organized, abstractions make sense, comments are helpful so I have a solid ground to do some good old organic human coding. I make sure that when i’m approaching limits I’m asking the AI to set the stage.

I used to get frustrated when credits ran out because the AI was making something I would need to study to comprehend. Now I’m eager to the next “brain time hand-out”

It sounds weird but it’s a form of teamwork. I have the means to pay for a larger plan but i’d rather keep my brain active.

> Don’t abuse DRY, a little duplication is better than unnecessary dependencies.

That's an interesting thing to include. I agree with this point in principle, but I've found that Claude, at least, duplicates logic FAR too often and needs nudging in the other direction.

> but I've found that Claude, at least

You hit on a very important point here. The linked AGENTS.md is a bad idea for general purpose use because the things it's meant to tackle, including an inherent bias towards or against DRY, is one of the big differences between model families. GPT 5.4 Codex has a very different "coding personality" from Claude Opus.

It's a product of whatever model it was tested on.

> It sounds weird but it’s a form of teamwork.

I can't do it. If I let an LLM write code for me, that code is untouchable. I see it as a black box, that I will categorically refuse to open. If it works, I use it, but don't trust it. If it breaks, I get frustrated. The only way that works for me is me behind the driving wheel at all times and an LLM as an assistant that answers my questions. We either brainstorm something or it helps me express things I know in languages syntax. Somehow that step has always been a bit of a burden for me - I understood the concepts well, but expressing them in syntax was a bit of a difficulty.

but demanding to be behind the wheel and understand all the code will affect velocity compared to other teams that are utilizing AI to the max

Well yes, if a team doesn't bother to understand the code, that's certainly quicker.

"Yes they crashed into a wall and all died, whereas you steered around it, but you must acknowledge that they crashed twice as quickly as you didn't crash. If you were driving their car, you would have just slowed them down."

People should have read to the end of "Building a C compiler with a team of parallel Claudes"[1]:

  The resulting compiler has nearly reached the limits of Opus [4.6]’s abilities. I tried (hard!) to fix several of the above limitations but wasn’t fully successful. New features and bugfixes frequently broke existing functionality.

"tried (hard!)" is very ominous. I wonder how Mythos would fare. Presumably it would get further, maybe much further. But I strongly doubt the "frequently broke existing functionality" problem was solved. Eventually humans have to understand the most difficult parts of the code. Good luck with that!

[1] https://www.anthropic.com/engineering/building-c-compiler

What good is "velocity" if no-one understands the direction we are going and how we are going? Sounds like a recipe for disaster.

Thanks for sharing. I have thought about approaches by deliberately leaving tasks to me while the agent does something to keep my brain active & prevent atrophy. Maybe I should work on a Claude Code skill/hook for that :)

Reply on news.ycombinator.com